OpenAI has released GPT-6 Astra, and it is not a subtle upgrade. It is the first model OpenAI is willing to describe with the phrase "AGI era," and it is also the most expensive model the company has ever put on its public API. Astra posts benchmark scores that were considered out of reach a year ago, and prices itself accordingly.
The headline is capability. GPT-6 Astra effectively saturates several of the hardest public benchmarks: 97.6% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and 100% on ExploitBench. It also sets a new bar on agentic computer-and-browser use, scoring 72.6% on OSWorld 2.0 while running roughly 47% faster per task than its predecessor, GPT-5.6 Sol. It has a 1.05-million-token context, 128K max output, and takes text and image input, with a knowledge cutoff of April 30, 2026.
The price of the frontier
Astra lists at $10 per million input tokens and $50 per million output, about 2.5x GPT-5.6 Sol's rate, and the priciest OpenAI has ever charged. The one relief is caching: cached input reads drop to $1, and for workloads that reuse context, that discount does most of the work in making Astra affordable at all. It's rolling out in stages, with the most cyber-sensitive capabilities gated behind a trusted-access program.
What it means
Two things. First, the frontier is now genuinely expensive again, after a year of prices falling, the best model costs real money, which pushes most everyday work toward the cheaper tiers (GPT-5.6, or rivals). Second, "AGI era" is a marketing phrase, not a technical milestone, and benchmark saturation isn't the same as reliability on your actual task. Astra is the most capable model available today; whether it's the right one for a given job still comes down to cost and fit. We track how it compares in the best AI models ranking.
